Robust Policy Gradient against Strong Data Corruption

Xuezhou Zhang, Yiding Chen, Xiaojin Zhu, Wen Sun
Proceedings of the 38th International Conference on Machine Learning, PMLR 139:12391-12401, 2021.

Abstract

We study the problem of robust reinforcement learning under adversarial corruption on both rewards and transitions. Our attack model assumes an \textit{adaptive} adversary who can arbitrarily corrupt the reward and transition at every step within an episode, for at most $\epsilon$-fraction of the learning episodes. Our attack model is strictly stronger than those considered in prior works. Our first result shows that no algorithm can find a better than $O(\epsilon)$-optimal policy under our attack model. Next, we show that surprisingly the natural policy gradient (NPG) method retains a natural robustness property if the reward corruption is bounded, and can find an $O(\sqrt{\epsilon})$-optimal policy. Consequently, we develop a Filtered Policy Gradient (FPG) algorithm that can tolerate even unbounded reward corruption and can find an $O(\epsilon^{1/4})$-optimal policy. We emphasize that FPG is the first that can achieve a meaningful learning guarantee when a constant fraction of episodes are corrupted. Complimentary to the theoretical results, we show that a neural implementation of FPG achieves strong robust learning performance on the MuJoCo continuous control benchmarks.

Cite this Paper

BibTeX
@InProceedings{pmlr-v139-zhang21d, title = {Robust Policy Gradient against Strong Data Corruption}, author = {Zhang, Xuezhou and Chen, Yiding and Zhu, Xiaojin and Sun, Wen}, booktitle = {Proceedings of the 38th International Conference on Machine Learning}, pages = {12391--12401}, year = {2021}, editor = {Meila, Marina and Zhang, Tong}, volume = {139}, series = {Proceedings of Machine Learning Research}, month = {18--24 Jul}, publisher = {PMLR}, pdf = {http://proceedings.mlr.press/v139/zhang21d/zhang21d.pdf}, url = {https://proceedings.mlr.press/v139/zhang21d.html}, abstract = {We study the problem of robust reinforcement learning under adversarial corruption on both rewards and transitions. Our attack model assumes an \textit{adaptive} adversary who can arbitrarily corrupt the reward and transition at every step within an episode, for at most $\epsilon$-fraction of the learning episodes. Our attack model is strictly stronger than those considered in prior works. Our first result shows that no algorithm can find a better than $O(\epsilon)$-optimal policy under our attack model. Next, we show that surprisingly the natural policy gradient (NPG) method retains a natural robustness property if the reward corruption is bounded, and can find an $O(\sqrt{\epsilon})$-optimal policy. Consequently, we develop a Filtered Policy Gradient (FPG) algorithm that can tolerate even unbounded reward corruption and can find an $O(\epsilon^{1/4})$-optimal policy. We emphasize that FPG is the first that can achieve a meaningful learning guarantee when a constant fraction of episodes are corrupted. Complimentary to the theoretical results, we show that a neural implementation of FPG achieves strong robust learning performance on the MuJoCo continuous control benchmarks.} }
Endnote
%0 Conference Paper %T Robust Policy Gradient against Strong Data Corruption %A Xuezhou Zhang %A Yiding Chen %A Xiaojin Zhu %A Wen Sun %B Proceedings of the 38th International Conference on Machine Learning %C Proceedings of Machine Learning Research %D 2021 %E Marina Meila %E Tong Zhang %F pmlr-v139-zhang21d %I PMLR %P 12391--12401 %U https://proceedings.mlr.press/v139/zhang21d.html %V 139 %X We study the problem of robust reinforcement learning under adversarial corruption on both rewards and transitions. Our attack model assumes an \textit{adaptive} adversary who can arbitrarily corrupt the reward and transition at every step within an episode, for at most $\epsilon$-fraction of the learning episodes. Our attack model is strictly stronger than those considered in prior works. Our first result shows that no algorithm can find a better than $O(\epsilon)$-optimal policy under our attack model. Next, we show that surprisingly the natural policy gradient (NPG) method retains a natural robustness property if the reward corruption is bounded, and can find an $O(\sqrt{\epsilon})$-optimal policy. Consequently, we develop a Filtered Policy Gradient (FPG) algorithm that can tolerate even unbounded reward corruption and can find an $O(\epsilon^{1/4})$-optimal policy. We emphasize that FPG is the first that can achieve a meaningful learning guarantee when a constant fraction of episodes are corrupted. Complimentary to the theoretical results, we show that a neural implementation of FPG achieves strong robust learning performance on the MuJoCo continuous control benchmarks.
APA
Zhang, X., Chen, Y., Zhu, X. & Sun, W.. (2021). Robust Policy Gradient against Strong Data Corruption. Proceedings of the 38th International Conference on Machine Learning, in Proceedings of Machine Learning Research 139:12391-12401 Available from https://proceedings.mlr.press/v139/zhang21d.html.

Related Material